Applying an ontology-aware zero-shot LLM prompting approach for information extraction in Greek: the case of DIAVGEIA gov gr
Dimitris Zeginis, Evangelos Kalampokis, Konstantinos A. Tarabanis · 2024
Large Language Models (LLMs) have attracted considerable attention, primarily due to their potential to revolutionize sectors that heavily rely on textual information.Governance is one such sector.Public administrations around the globe produce millions of documents including laws, administrative decisions and acts (e.g., travel/budget approvals) that contain valuable information in unstructured way.The documents are usually stored at documentcentered repositories.As a result the actual data of the documents cannot be further searched or processed.The availability of structured metadata of the documents (e.g., who has traveled, where, when) could further enhance the searching and processing of the documents as well as enable data analytics.The construction of metadata can be done through information extraction approaches such as Named Entity Recognition (NER), Relation Extraction (RE) and Event Extraction (EE) on the documents.LLMs are recently used successfully for information extraction tasks, while ontologies are traditionally used for meaningful data modeling.The aim of the paper is to apply and evaluate an ontology-aware zero-shot LLM prompting approach for information extraction in Greek language documents available in DIAVGEIA.gov.gr-the Greek Open Government portal for administrative documents.The evaluation assesses various LLM models/sizes for various difficulties of information extraction tasks.Overall the results are very promising, since most LLM models, even smaller ones, performed very well for all tasks in Greek.